mcp-security-audit

MCP Security Audit: The Sentinel Your MCP Configs Deserve

The MCP Security Audit skill earns its place in the CLI-and-API cluster by addressing a pain point that's both specific and urgent: securing your Model Context Protocol server configurations. While other skills in this cluster focus on automation, data fetching, or tool orchestration, this one takes on the thankless but critical job of governance. It doesn't just read your .mcp.json files—it interrogates them for secrets, shell injection vectors, and supply-chain risks. That focus alone makes it worth a second look.

What sets it apart is its laser-like scope. The skill doesn't try to be a general-purpose security scanner; it's built for one config file and one threat model. This precision means it can deliver actionable results without noise. In testing, the list-or-read test returned a structured JSON report with clear findings—no fluff, no false positives. It flagged hardcoded credentials and unpinned dependencies exactly where you'd expect them. The install-and-auth test passed cleanly, confirming the skill is a self-contained Python CLI tool with no external API dependencies. That's a rare treat for a security tool.

Where it stumbles is the write-or-mutate test, which failed because the skill is read-only. There's no --fix flag, no auto-remediation. This is by design—the skill audits, it doesn't fix—but in practice, you'll need to pair it with a separate remediation step. The reusability score of 3.5 reflects this limitation: it's excellent for one-shot audits but less useful in CI/CD pipelines where you want automated fixes.

Two observations from testing stand out. First, the skill correctly identified shell injection patterns in MCP server arguments, something that's easy to miss in manual review. Second, it validated pinned version requirements against @latest tags, catching a dependency that was floating. Both findings were precise and clearly documented in the JSON output.

When would you reach for this skill? In any project where you're onboarding MCP servers or reviewing a teammate's configuration. It's especially valuable during code review or before a deployment. Pair it with a mutation-capable skill for remediation, and you've got a solid security checkpoint. For the audit itself, this skill is the right tool—focused, fast, and honest about its limits.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-07-24.

Overall: partial. 3 tests passed, 0 partial, 1 failed; key blocker: write/mutate not supported (read-only audit).

Inferred dependencies: python>=3.10, pyyaml, jsonschema.